To Copy or Not to Copy: Copying Is Easier to Induce Than Recall

2026-01-17

AI Summary

This study investigates the arbitration mechanism between parametric knowledge (recall) and retrieved context (copying) in retrieval-augmented language models. By constructing a dataset that distinguishes irrelevant contexts (eliciting superficial recall) from relevant but incorrect ones (inducing erroneous copying), the authors propose an arbitration vector derived from residual stream centroid differences and inject it into model activations to directionally modulate behavior. The work reveals, for the first time, an asymmetry between copying and recall: copying is a robust reactivation process, whereas recall is a fragile inhibitory process dependent on intervention at specific token positions. Combining attention routing, MLP contribution analysis, and layer-wise probability trajectories, the approach achieves consistent and controllable behavioral switching across two architectures and two open-domain question answering benchmarks, maintaining both accuracy and generation fluency.

Technology Category

• Data Mining & Knowledge Management: Conversational Systems for Recommendation & Retrieval
• Natural Language Processing: Discourse, Pragmatics & Argument Mining
• Machine Learning: Calibration & Uncertainty Quantification

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Language models used in retrieval-augmented settings must arbitrate between parametric knowledge stored in their weights and contextual information in the prompt. This work presents a mechanistic study of that choice by extracting an \emph{arbitration vector} from model activations on a curated dataset designed to disentangle (i) irrelevant contexts that elicit parametric recall and (ii) relevant but false contexts that elicit copying. The vector is computed as the residual-stream centroid difference between these regimes across 27 relations, and is injected as an additive intervention at selected layers and token spans to steer behavior in two directions: Copy$\rightarrow$Recall (suppressing context use) and Recall$\rightarrow$Copy (inducing the model to copy any token from the context). Experiments on two architectures (decoder-only and encoder/decoder) and two open-domain QA benchmarks show consistent behavior shifts under moderate scaling while monitoring accuracy and fluency. Mechanistic analyses of attention routing, MLP contributions, and layer-wise probability trajectories reveal an asymmetry: inducing copying is an easy ``reactivation''process that can be triggered at different locations in the input, while restoring recall is a ``suppression''process that is more fragile and strongly tied to object-token interventions.

Problem

• retrieval-augmented generation
• parametric knowledge
• contextual information
• copying vs. recall
• language models

Innovation

• arbitration vector
• retrieval-augmented generation
• mechanistic interpretability
• copying vs. recall
• activation intervention